metadata
license: other
license_name: tongyi-qianwen
license_link: https://huggingface.co/Qwen/Qwen2-72B-Instruct/blob/main/LICENSE
pipeline_tag: text-generation
language:
- en
- zh
library_name: transformers
tags:
- mergekit
- llama
FeatherLlama-72B-v0.1
Model Description
FeatherLlama is a 72B parameter language model created through a merge of Qwen2-72B-Instruct, calme2.1-72b, and magnum-72b-v1 using model_stock.
This is converted from leafspark/FeatherQwen2-72B-v0.1
Features
- 72 billion parameters
- Sharded in 31 files (unlike FeatherQwen2, which has 1,043 shards due to the merging process)
- Combines Magnum prose with Calam smarts
- Llamaified for easy use
Technical Specifications
Architecture
LlamaForCasualLM- Models: Qwen2-72B-Instruct (base), calme2.1-72b, magnum-72b-v1
- Merged layers: 80
- Total tensors: 1,043
Tensor Distribution
- Attention layers: 560 files
- MLP layers: 240 files
- Layer norms: 160 files
- Miscellaneous (embeddings, output): 83 files
Merging
Custom script utilizing safetensors library.
Usage
Loading the Model
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
model = AutoModelForCausalLM.from_pretrained("leafspark/FeatherLlama-72B-v0.1",
device_map="auto",
torch_dtype=torch.float16)
tokenizer = AutoTokenizer.from_pretrained("leafspark/FeatherLlama-72B-v0.1")
GGUFs
Find them here: leafspark/FeatherLlama-72B-v0.1-GGUF
Hardware Requirements
- Minimum ~140GB of storage
- ~140GB VRAM